基于Mask R-CNN的马匹烙印号提取方法研究及应用

Research and application of horse brand number extraction method based on Mask R-CNN

马匹烙印号提取是鉴定工作不可缺少的环节。针对传统的鉴定工作需人工进行标记,且存在工作量大、效率低、人工成本较高等问题,采用了Mask R-CNN算法对马匹烙印号进行分割提取。通过Labelme标注出马匹躯干及烙印号制作数据集标签,结合主干网络(ResNet-101-FPN)进行特征提取,区域建议网络(RPN)生成感兴趣区域,采用ROI-Align层,将感兴趣区域(ROI)的特征池化为固定尺寸的特征图。由检测分支进行目标框的分类和回归,实现图像背景与特征区域的分割。实验结果表明,使用实例分割算法检测马匹烙印号的准确率达到0.998,分割的平均精度(Average Precision,AP)是0.944,在准确检测的同时实现了像素级的马匹躯干及烙印号信息分割,可为马匹鉴定工作提供便利。

Horse brand number extraction is an indispensable link in the process of horse identification.When traditional identification methods are used,horses need to be marked manually.This paper uses Mask R-CNN algorithm to segment and extract horse mark numbers so as to solve the problems of high labor cost and low work efficiency.Labelme is used to mark horse body and mark numbers which will be then used to make dataset labels.Features are used to extract backbone network(ResNet-101-FPN)and RPN so as to generate their regions of interest.ROI-Align layer is used to pool features of the ROI to fixed dimensional features.The classification and the regression of target frames are then realized through detection branch es so as to segmentimage background and feature regions.The experimental results show that the instance segmentation algorithm can achieve a high accuracy rate of horse brand mark detection,which is 0.998.The Average Precision(AP)is 0.944.This algorithm can realize accurate detection and segmentthe information of horse trunk and brand number can being at pixel level,which can provide convenience for horse identification.